What's new in the picker since September

Six of the newest models in the tryingopen picker arrived in September and October. Five of them read pictures here, two run with zero retention by default (Ling 3.0 Flash VL and LFM2.5 24B), and Mistral Large 4's weights are still to come. Everything below is as of October 14, 2026.

"Provider may log" comes up a lot below. It is the picker's label for a model whose host might keep what you send. What zero data retention covers, and where it stops explains what the label means for your messages.

The six new models in one table

DeepSeek V4.1 Flash

DeepSeek released V4.1 Flash on September 10, 2026, and we added it on September 20. Its API changelog calls it "the smallest model in our new architecture family, with native multimodal visual understanding". The model card puts the weights under the MIT license, with up to 1M tokens of context and image and text input.

Per the card, it activates 8B parameters per token at prefill (while it reads your prompt) and 16B at decode (while it writes the reply). Its total size depends on what you count, so we leave that to What total and active parameters tell you about a model.

On DeepSeek's own API, V4.1 Flash replaced V4 Flash. The changelog says the older V4 Flash models "have been retired" and that their names are "temporarily routed to V4.1 Flash." Tryingopen still offers DeepSeek V4 Flash as a separate pick, so you can ask both the same question in two chats.

On tryingopen, V4.1 Flash reads pictures. With no provider picked, the composer shows it as run by QwenCloud and labels it "Provider may log". Pick a provider in the composer and the message goes through OpenRouter with a zero-retention requirement instead. It is also the model tryingbaseten.com starts on, the version of the site where Baseten runs every message. Choosing the company that runs your model covers the provider picker.

Xiaomi MiMo V2.6 Flash and Pro

OpenRouter listed Xiaomi's MiMo V2.6 models on September 21, 2026 (UTC), with the weights on Hugging Face under MIT, and we added both on September 22. The Flash card gives 309B total parameters with 15B active and a 1M-token context. The picker shows no size for Pro. OpenRouter's listing says only "over 1T parameters", though Xiaomi's Pro card gives 1.02T total with 42B active.

Both read pictures on tryingopen, and both carry "Provider may log". With no provider picked, the composer shows them as run by Tencent TokenHub, reached through ngrok's gateway. Neither has a provider to pick: when we tested them on September 22, Xiaomi was the only host on OpenRouter, and a request pinned to zero retention found no endpoint.

Self Chat uses only zero-retention routes, so it leaves both MiMo models out (Self Chat saves your chats only where you choose).

Ling 3.0 Flash VL

Ling 3.0 Flash VL comes from inclusionAI, Ant Group's AI lab. It joined the picker on September 22, the same day as the MiMo models. Its model card gives 124B total parameters with 5.5B active per token, image and video input, and the MIT license.

On tryingopen it reads pictures and is zero-retention by default, with no provider pick needed. Self Chat offers it too.

Mistral Large 4

On October 6, 2026, Mistral announced Mistral Large 4 as a public preview API, describing "a 1 trillion-parameter natively multimodal model with 52 billion active parameters". We added it on October 8. It reads pictures here.

It carries two labels in the picker:

• Weights not out yet. Mistral's announcement says "We will release the weights by the end of the month." As of October 14, Mistral's Hugging Face page has no Large 4 weights.
• Provider may log. It has no zero-retention route today, so Self Chat leaves it out as well.

The missing weights keep it out of the picker's count of 23 open models. What counts as an open model, and what the licenses allow covers what "open" means on this site.

d1 and LFM2.5 24B from Liquid AI

• d1, released October 7: Liquid AI's open decision model. Since October 12 it has decided, for each message sent to the default pick, whether our small model or a cloud model you choose answers it. See How the d1 router picks between our small model and the cloud.
• LFM2.5 24B, released October 13: a text-only, zero-retention pick that arrived the same day. See LFM2.5 24B joined the picker the day Liquid AI released it.

What we test before a model goes in

We send it a request pinned to zero retention through OpenRouter. Other test requests check whether it reads pictures and calls tools as its OpenRouter listing says. The MiMo models and Mistral Large 4 failed the zero-retention check and went in anyway, labelled "Provider may log". The picture test can switch a feature off. OpenRouter lists Muse Glimmer 30B as taking images, but in our test it answered as if no picture was attached, so picture upload is off for it here.

Also new since September

Two more models joined the picker for the first time: Nemotron 3 Ultra on September 20 and Tencent's Hy4 Preview on October 8. Three that left the picker in August, GLM 5.3, Inkling and Qwen3.8 2.4T, came back on September 20. Two picks run on our own hardware: Liquid AI's LFM2.5-VL 3B became the default on September 30 (Our default model runs on a palm-sized Jetson Orin Nano explains it), and Gemma 4 E2B joined on October 2 (Gemma 4 on tryingopen, from our boards to the cloud).

Three of the models above (DeepSeek V4.1 Flash and both MiMo models) are also on the published model list for the developer API at tryingopen.com/developers. The API docs are public. As of October 14, 2026, developer accounts are not open yet.

A tour of the 24 models in the picker covers the rest of the list.